franecki/pill_bottle_203

SO-ARM101 Pill Bottle Pick-and-Place Dataset (203 Episodes)

Task: Pick up a white pill bottle (cylinder) and place it inside a black electrical-tape circle.

Prompt (single task): Pick up the white pill bottle and place it inside the black tape circle.

Overview

Property	Value
Robot	SO-ARM101 (6-DoF follower: 5 joints + gripper)
Control	Teleoperation (leader arm)
Episodes	203
Total frames	77,944
FPS	30
Total duration	~43 minutes
Dataset version	LeRobot v3.0
Codec	AV1 (libsvtav1 / libdav1d), YUV420p
Camera resolution	640 x 480
Split	train = episodes 0-202

Data Format

LeRobot v3 dataset layout:

pill_bottle_203/
├── meta/
│   ├── info.json          # dataset metadata (features, fps, splits, codec)
│   ├── episodes           # per-episode metadata (video file + timestamp ranges)
│   ├── stats.json         # normalization statistics per feature
│   └── tasks.parquet      # task strings
├── data/
│   └── chunk-000/         # parquet files: file-000..file-NNN.parquet
├── videos/
│   ├── observation.images.top/
│   │   └── chunk-000/     # AV1 video files
│   └── observation.images.gripper/
│       └── chunk-000/
└── README.md

Features (info.json)

Key	Type	Shape	Content
observation.images.top	video (AV1)	480x640x3 RGB	top-down camera view of workspace
observation.images.gripper	video (AV1)	480x640x3 RGB	gripper-side close-up camera
observation.state	float32	[6]	joint positions (normalized -100..100)
action	float32	[6]	joint target positions (normalized -100..100)
timestamp	float32	[1]	time within episode (s)
frame_index	int64	[1]	global frame index
episode_index	int64	[1]	episode id (0-202)
index	int64	[1]	global row index
task_index	int64	[1]	task id (0)

Joint order (observation.state / action)

[shoulder_pan, shoulder_lift, elbow_flex, wrist_flex, wrist_roll, gripper]

• Positions are normalized to the calibrated motor range, units span roughly -100..+100 (RANGE_M100_100), gripper uses RANGE_0_100 (open ≈ low, close ≈ higher).
• use_degrees = False (normalized range, not radians/degrees).

Recording Setup

• Follower (slave) arm on /dev/ttyACM1, calibrated id my_awesome_follower_arm.
• Leader (master) arm on /dev/ttyACM0, calibrated id my_awesome_leader_arm.
• Cameras: top = /dev/video0, gripper = /dev/video2, both MJPG 640x480 @30fps, RGB.
• Every episode starts from the folded home pose: shoulder_pan ≈ -3.5, shoulder_lift ≈ -98.5, elbow ≈ +98.7, wrist_flex ≈ -98, wrist_roll ≈ -50, gripper ≈ 2 (open).
• Object placement (bottle and tape circle) is randomized across the workspace center region per episode.

Notes on a few episodes

• Episodes 31, 136, 198 start with a slightly different home pose (wrist/gripper off by a few units from the standard folded pose). These are valid demonstrations; the model conditions on the observed state so this introduces only minor start-state variation.

How to Load

from lerobot.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("franecki/pill_bottle_203")
episode = ds[0]  # returns frames with observation.images.*, observation.state, action

Or via CLI after upload:

huggingface-cli download franecki/pill_bottle_203 --repo-type dataset --local-dir ~/.cache/huggingface/lerobot/franecki/pill_bottle_203

Notes

• Videos are AV1 encoded. Decoding requires torchcodec or pyav/libdav1d (OpenCV alone cannot decode AV1).
• The observation.state and action values are already normalized to the calibrated motor range; use meta/stats.json (already computed) for the training pipeline's normalization stats.
• Video files may contain a small number of extra (orphan) frames at episode boundaries that are never referenced; they are ignored by timestamp-based frame retrieval.